Vulnerability disclosure
If you’ve found a security issue in Carabase Host — credential leakage, authentication bypass, RLS escape, supply-chain risk, anything else that breaks the threat model — we want to hear about it before it’s public.
How to report
Section titled “How to report”Email security@carabase.dev. Send your report privately so we can triage and fix before any public disclosure.
Please include:
- A description of the issue and what an attacker could do with it
- Affected version(s) —
GET /api/v1/versionreports the running version - Steps to reproduce, or a proof-of-concept if you have one
- Your preferred attribution (real name / handle / anonymous)
What we promise
Section titled “What we promise”- Acknowledgement within 72 hours of receiving the report
- A fix or mitigation plan within 14 days for high/critical issues, 30 days for moderate, best-effort for low
- A coordinated disclosure window — we’ll agree on a date for public disclosure together, default 90 days from acknowledgement
- Public credit in the security advisory, unless you prefer to stay anonymous
In scope:
- The Carabase Host application at any tagged release
- The Admin SPA bundle served at
/admin/ - The inbound MCP surface (
/mcpstreamable HTTP and the legacy/mcp/sse+/mcp/messagesSSE transport) and its access gate - The operator shell scripts that ship with the host
Out of scope:
- The OpenClaw gateway itself — that’s a separate project with its own disclosure process
- The local vendor CLIs that the
codex_cli/claude_codechat and agent runtimes shell out to (Codex, Claude Code) — report those to their respective vendors; issues in how the host invokes them (e.g. an escape past the read-only token gate) are in scope - Tailscale — report directly to tailscale.com/security
- Issues in third-party connectors (GitHub, Google, Granola, and the other integrations) — report to the respective vendor
- Dependency CVEs that already have a public CVE — report it to us by email
- DoS / resource exhaustion attacks — single-tenant, self-hosted; denying the service to yourself isn’t a meaningful attack
- Social engineering / phishing the maintainer — not a software vulnerability
What you should NOT do
Section titled “What you should NOT do”- Don’t test against installs you don’t own
- Don’t publicly disclose before the agreed window
- Don’t demand bug bounty payment — this is a small project run by individuals, not a company with a budget. We’ll do our best to credit and amplify your work, but we can’t pay
Recent advisories
Section titled “Recent advisories”(None yet. This page will list all advisories once we publish any.)